Turbo King: Framework for Large- Scale Internet Delay Measurements Derek Leonard Joint work with Dmitri Loguinov Internet Research Lab Department of Computer Science Texas A&M University, College Station, TX 77843 April 15, 2008 1
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 2
Introduction I Distance estimation in the Internet has recently evolved into a large field The goal is to estimate or measure latency (delay) between hosts Can be used to provide better service to endusers and construct more efficient networks Increasing responsiveness for online games Quickly locating the closest server in a CDN Creating topologically-aware P2P networks 3
Introduction II Focus of our work Create a method for distance estimation that requires no infrastructure to be deployed throughout the Internet Allow for the generation of a much larger Internet latency matrix than previous work Requirements Produce accurate latency estimates Minimize the impact of our measurements on the network 4
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 5
King I King uses existing DNS infrastructure to estimate the latency between two arbitrary hosts in the Internet Assumes end-hosts are within close proximity to the authoritative nameserver responsible for their IP addresses 6
King II Original King (O-King) Main algorithm proposed by Gummadi et al. Uses queries for authoritative zone data to measure the latency between two remote DNS nameservers Direct King (D-King) Also proposed by Gummadi et al., but not fully implemented or evaluated in the literature Alternative to O-King that forces a nameserver to query an arbitrary target server 7
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 8
Understanding O-King I While O-King s simplicity is attractive, it has certain drawbacks Zones with multiple authoritative nameservers Recommended by DNS RFC Suggested to be placed in different networks 33% of zones have at least one server in a different network 9
Understanding O-King II DNS Forwarders Server that aggregates DNS queries initiated from within a network targeting external destinations Recursive nameservers receiving requests for zones not under their control often use forwarders to reduce their own load End-user not notified Unnoticed by O-King Potentially introduces non-trivial extra latency 10
Understanding O-King III Cache pollution: insertion of DNS zone data that has not been requested by a local user into the cache of a nameserver The purpose of a DNS cache is to reduce latency mainly for local users Local users are those that rely principally on the nameserver to resolve queries O-King seed queries pollute the cache with two records for each target nameserver At large scale records inserted by O-King would dominate local caches Would likely be viewed as intrusive by admins 11
Understanding D-King I Additional Complexity A domain name and extra infrastructure (DNS server with dedicated IP) are required O-King is so simple, is D-King worth it? Forwarders D-King does not detect or avoid forwarders Cache Pollution D-King only requires caching records for one server However, the cached records are completely useless to others 12
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 13
Turbo King I Turbo King (T-King) basics Stand-alone service Accepts as arguments the IP addresses of two end-hosts A and B Returns estimated latency from host A to B Two modes of operation Passive: waits for requests before generating latency estimates Active: preemptively make latency estimates to eventually obtain an entire N Nmatrix 14
Turbo King II Server selection is a major difference between King and Turbo King Turbo King maintains a large list S of N nameservers Currently both recursive nameservers and other authoritative nameservers Use BGP data to match servers from S to both A and B Negates the assumption that the authoritative nameservers for end-host IP addresses are closest 15
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 16
Measurement Algorithm I Measurement algorithm Multi-threaded application with client and server operations communicating together Timestamps taken for every packet sent/received by T-King All answers returned by DNSServer have zero TTL Let d ij be the delay between steps i and j 17
Measurement Algorithm II Local latency sample L i = d 12 Remote latency sample R i = d 36 Latency estimate: min{r i } -min{l i } Detection and avoidance of forwarders Compares IP addresses used to contact T-King If different, exclude the original query IP and add the newly discovered one to S for later use 18
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 19
Discovering Nameservers I Discovering Nameservers T-King is most effective when its list S of nameservers is large Ideally S would contain a nameserver for each /24 network in the Internet The current version builds S by exhaustively crawling the reverse DNS (IP to domain name) tree Only accepts authoritative responses Maximizes depth of the crawl and subsequently the number of discovered nameservers 20
Discovering Nameservers II Results from the reverse DNS crawl Our crawler is approximately seven times faster than a previous effort and discovered 2.4 times more nameservers We found that 32% use a forwarder for queries 21
Discovering Nameservers III Coverage of Internet by discovered servers We found that 49% of Gnutella peers and 88% of web servers are in a BGP prefix that contains at least one nameserver in T-King For recursive servers, 37% of Gnutella and 73% of web servers are covered 22
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 23
Evaluation I We took 2,450 measurements for 50 nameservers using O- King and T-King Ratio of O-King to T-King used to highlight the differences Four samples per estimate Conclusion: 15% of O-King estimates are more than 10% different from T-King and 8% of O-King estimates are more than 20% different 24
Evaluation II We next compare the convergence properties of T-King vs. O-King Consistent samples are the goal so that an accurate estimate requires fewer samples To do this, we repeated the previous 2,450 latency estimates using sample sizes ranging from one to four We then calculated the ratio of O-King to O-King and T-King to T-King and plotted the CDF More consistent samples are centered at one on the x-axis 25
Evaluation III Conclusion: The original suggestion that O- King use 4 samples is sound Conclusion: T-King converges to a consistent estimate in only 2 samples 26
Evaluation IV The overhead associated with each method is critical when considering a large-scale measurement Consider the generation of a hypothetical latency matrix of 100,000 100,000 hosts Requires 10 billion estimates to complete Consider 4 samples per estimate for O-King, 2 samples for both D-King and T-King Examine both total number of queries sent and total cache pollution entries 27
Evaluation V Network overhead to complete 10 billion estimates Turbo King: 70 billion queries D-King: 100 billion queries (1.43 times T-King) O-King: 150 billion queries (2.14 times T-King) Conclusion: The increased accuracy and lack of seeding required by T-King results in a significant reduction in bandwidth usage 28
Evaluation VI Cache pollution created for 10 billion estimates Two entries for every nameserver entered into cache Total polluted entries in DNS caches O-King: 48 billion entries (2.4 nameservers per zone measured from our reverse crawl) D-King: 20 billion entries T-King: 200,000 entries (0.0004% of O-King and 0.0001% of D-King) Conclusion: T-King is much more suitable for large-scale measurements in this regard 29
Agenda Introduction Related Work Understanding King Turbo King Measurement Algorithm Discovering Nameservers Evaluation Conclusion 30
Conclusion We proposed Turbo King as a framework to perform large-scale Internet latency measurements More accurate than both O-King and D-King Requires fewer samples than O-King Much more scalable in terms of both bandwidth usage and cache pollution Our next step is to generate an N Nlatency matrix using T-King Please see the paper for more details 31